Embedding watermarks into deep neural networks of audio classification

• Main Author: Chin, Jun Ying
• Other Authors: Zhang Tianwei
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering; Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/147918
• Institution: Nanyang Technological University

In recent years, there is an increasing trend of developing high performance neural network to tackle various real-world problems. This has led to momentous progress areas such as image recognition, speech emotion analysis and natural language processing. Significant amount of training data, computer resources and human resources are required to produce a service-grade neural network. Hence, it is important to regard and protect neural networks as intellectual property, owned by the creators. Various digital watermarking techniques have been proposed to identify violation of intellection property of such networks, primarily neural networks build for image classification problems. This project focuses on investigating the effectiveness of backdoor-based watermarking techniques on neural networks dealing with audio classification, then investigates the effectiveness of three different watermark generation algorithms. Additional techniques that enhance the robustness of watermarks embeddings are also explored. These include making the watermark embedding resistant against typical transformations of data in the audio domain, pruning, and fine-tuning of the trained model. This project ultimately aims to identify an effective method of watermarking of neural networks in the audio domain.